Equity

Max Spero on building the internet's trust layer

Max Spero· Co-founder and CEO of Pangram at Pangram
·~24 min·English·TechCrunch
AI SafetyMultimodalAI Company
TL;DR

Pangram co-founder and CEO Max Spero argues the internet needs a trust layer that measures the degree of AI in text and images — not a yes-or-no verdict — so platforms can find what is human and prioritize it as bots, slop, and AI-enabled fraud multiply.

01Why a Trust Layer

The Imbalance of Effort

<strong>AI makes producing content nearly free while checking it stays expensive</strong> — one person can now fire off a thousand job applications a day, and the human reviewer on the other side drowns.

And so this creates this just like massive imbalance of effort where it's infinitely easy to produce content or take actions and then the verifier or reviewer on the other side ends up being kind of swamped.

— Max Spero, Equity
Key Insight
The problem Pangram sells against is not that AI writes badly — it is that AI writes cheaply. When the cost of generating plausible text falls to zero, every system that quietly assumed human effort as a rate limit — hiring, reviews, claims — loses its throttle at once.
02The Product Bet

From Yes/No to a Dial

<strong>Pangram stopped asking whether a text is AI and started measuring how much AI it is</strong> — the useful question is the degree of assistance between a grammar fix and a fully generated draft.

We've went from like something that's kind of binary of like is there AI yes or no to really more like mixed trying to understand like what is the degree of AI assistance versus human input.

— Max Spero, Equity
Key Insight
A binary detector forces a false choice: accuse or clear. A dial reframes detection as disclosure — the point is not to ban AI help but to measure how much of a piece is still the author's. That is the only version of detection that survives a world where almost everyone uses some assistance.
03How It Works

A Thousand Weak Signals

<strong>No single word proves anything; the verdict is built from hundreds of tiny tells.</strong> Each word choice is a weak hint, and across 50 to 200 words those hints compound into high confidence.

then I think we can build up confidence over time through a bunch of weak signals to say that we're actually very confident that this is AI generated.

— Max Spero, Equity
Key Insight
This is why Pangram does not rely on watermarks. A watermark is one strong signal a model can strip; a thousand weak stylistic signals cannot all be scrubbed without rewriting the text into something human. The method trades a single point of failure for statistical weight.
04The Stakes

Where the Money Leaks Out

<strong>AI slop is merely annoying in a feed but expensive everywhere else</strong> — fake receipts, staged refund photos, bot reviews, and mass fraud all ride on content that is now free to fabricate.

I mean, I think we've just seen an increase in fraud that's enabled by AI. So, anything from like submitting a fraudulent expense and then having AI like produce a receipt for you to like, you know, like having AI put a bug in your food and then sending it to Door Dash to get a refund.

— Max Spero, Equity
Key Insight
The consumer framing — 'I don't want to see slop' — undersells the business case. Spero's real market is the back office: the expense system, the refund queue, the review platform, each of which assumed forged evidence was hard to make. Detection becomes fraud infrastructure, not a content filter.
05The Image Frontier

You Can't Count the Fingers Anymore

<strong>The old tells for fake images are gone.</strong> Six fingers and garbled text used to give AI images away; today's generators produce photos Spero says he cannot tell apart from real ones.

It's because AI images have gotten really good kind of scarily fast. I think we went from AI images being obvious. You just count the fingers. You look at the text and make sure it's not garbled.

— Max Spero, Equity
Key Insight
When the human eye loses the ability to spot a fake, verification has to move from perception to statistics — the same weak-signal approach Pangram uses on text. The deeper stake Spero names is reality itself: if any photo of the past can be silently regenerated, the historical record stops being evidence.
06The Hard Tradeoff

One in Ten Thousand Is Not Zero

<strong>A near-perfect detector still flags innocent people at scale.</strong> A 1-in-10,000 false-positive rate means dozens of wrongly accused humans a day, so Pangram leans on multiple pieces of evidence over a single verdict.

So, Pangram has a 1 in 10,000 false positive rate. That means text that was fully human written will get flagged as AI about 1 in 10,000 times, which is very much not zero.

— Max Spero, Equity
Key Insight
Spero's honesty about the false-positive rate is also a liability strategy. By publishing the number and pushing multiple accounts over single verdicts, Pangram tries to reposition itself from judge to evidence — the tool that raises healthy skepticism, not the one that convicts. Being called a 'defamation machine' in the press is exactly the risk he is managing.
07The Structural Shift

Saving Us From the Dead Internet

<strong>Spero thinks the dead-internet theory is close to real</strong> — and the response is a retreat into walled gardens, closed group chats, and an internet increasingly tied to real names and faces.

I think we're we're actually dangerously close if if we do nothing like dead internet theory will happen within the next few years.

— Max Spero, Equity
Key Insight
There is a tension Spero does not resolve: his product helps preserve the open internet by labeling bots, yet the future he predicts gives up on openness and hides behind identity walls. Detection may be less a cure for the dead internet than a stopgap before people wall themselves off anyway.
08The Human Premium

Slop Is the Volume, Not the Tool

<strong>Using AI is not what makes something slop — churning out low-quality work at high volume is.</strong> Where quantity beats quality, AI wins; where quality matters, the human voice still commands a premium.

I think slop and like AI content in general like is is associated with lowquality loweffort outputs. We're hiring writers. We're working with writers directly. Like everything that we put out is human written.

— Max Spero, Equity
Key Insight
Spero draws the line where it is commercially convenient for a detection company: AI is fine for commodity text, humans win on quality. But his own example cuts deeper — much SEO copywriting was 'human slop' before AI existed, which means the technology did not create low-value writing, it automated a category that was already hollow.